fix: Add performance numbers, Scan snapshot, and streaming rationale to README

[IsaacCheng9]

Summary

• Added a top-level ## Performance section with a 6-row table covering memtable / SSTable / write / gRPC numbers
  • Each row shows both throughput (e.g. 2.6M ops/sec memtable reads, ~117k rows/sec gRPC streaming scans) and p50 latency, derived from docs/2026_05_05_grpc_with_scan_baseline.txt
• Added a Scan Snapshot node to the architecture diagram
  • Snapshot semantics for Scan (consistent reads across concurrent writes, flushes, and compactions) was previously invisible in the data-flow picture, despite being the most distinctive Phase 2 engineering choice
  • Relabelled the existing Engine -> L0/L1 read arrows as 'point reads' to distinguish from the snapshot's new 'range reads', so the diagram no longer appears to show two redundant Engine -> SSTable paths
• Renamed the gRPC ### Performance subsection to ### Why Server-Streaming for Scan
  • Trimmed the now-redundant numbers paragraph (the same numbers are in the top-level table) so the section focuses on engineering rationale rather than re-stating data